Abstract:Time-series data are widely used in fields such as industrial manufacturing, meteorology, electric power, and vehicles, which has spurred the development of time-series database management systems. More and more database systems are migrating to the cloud, and the architecture of end-cloud collaboration is becoming more common, leading to increasingly large data scales to be processed. In scenarios such as end-cloud collaboration and massive time series, a large number of short time series are generated due to short synchronization cycles and frequent data flushing, among other reasons, presenting new challenges to database systems. Efficient data management and compression methods can significantly improve storage performance, enabling database systems to handle the storage of massive time series. Apache TsFile is a columnar storage file format specifically designed for time series scenarios, playing an important role in database management systems such as Apache IoTDB. This study elaborates on the group compression and merging methods used in Apache TsFile to address scenarios with a large number of short time series, especially in application scenarios with a vast number of time series such as the Industrial Internet of Things. This group compression method fully considers the data characteristics in the short time series scenario. Through device grouping, it improves metadata utilization, reduces file index size, decreases short time series, and significantly improves compression effectiveness. After validation with real-world datasets, the proposed grouping method shows significant improvements in compression effect, reading, writing, file merging, and other aspects, enabling better management of TsFiles in short time series scenarios.

Abstract:In recent years, with the development of software and hardware, migrating databases to the cloud has become an emerging development trend and can reduce database operation and maintenance costs for small and medium-sized enterprises and individual users. Furthermore, the development of cloud databases has led to a massive market demand for database operation and maintenance. Researchers have proposed many database self-tuning technologies to support automatic optimization of database knobs. To improve tuning efficiency, existing technologies have shifted from focusing solely on the tuning problem itself to focusing on how to reuse historical experience to find the optimal parameter configuration for the current database instance. However, with the development of cloud databases, users have gradually increased their requirements for privacy protection, hoping to avoid privacy leakage while having efficient data access efficiency. Existing methods do not consider protecting the privacy of users’ historical tuning experience, which may cause user load characteristics to be perceived, causing economic losses. This study analyzes the characteristics of cloud database tuning tasks in detail, organically combines the server side and the user side, and proposes a cloud database knob tuning technology based on federated learning. First, to solve data heterogeneity in federated learning, this study proposes an experience screening method based on meta-feature matching to eliminate historical experiences with large differences in data distribution in advance to improve the efficiency of federated learning. To protect user privacy, this study organically combines the characteristics of cloud database services and proposes a federated Bayesian optimization algorithm with the node end as the training center. Through random Fourier features, it achieves user privacy protection without distorting the tuning experience. The results on extensive public benchmarks present that the proposed method could achieve competitive tuning performance compared with existing tuning methods. Moreover, due to the reuse of historical experience, it can greatly improve tuning efficiency.

Abstract:One of the most important features of multi-tenant databases in cloud environments is scalability. However, most elastic scaling techniques struggle to make effective scaling decisions for dynamically changing loads. If load changes can be predicted in advance, resource supply can be accurately adjusted. Given this, this study proposes a load-prediction-based elastic scaling method for multi-tenant databases. It includes a combined load prediction model and an elastic scaling strategy. The load prediction model combines the advantages of convolutional neural networks, long short-term memory networks and gated recurrent units. It can accurately forecast memory requirements of database clusters. Based on the prediction results, the elastic scaling strategy adjusts the number of virtual machines to ensure that resource supply remains within a reasonable range. Compared to existing methods, the combined load prediction model can reduce prediction errors by 8.7% to 21.8% and improve prediction fitting degree by 4.6%. Furthermore, this study improves the Bayesian optimization algorithm for hyperparameter tuning of the combined prediction model. The improved hyperparameter tuning model reduces errors by above 20% and improves fitting degree by 1.04%, which proves that it can well address the poor performance of Bayesian optimization in combined domains of discrete and continuous solutions. Compared to the most widely used scaling strategy in Kubernetes, the proposed elastic scaling method reduces response time by 8.12% and latency by 9.56%. It can avoid the latency and the waste of resources to a large extent.

Abstract:Cloud-native databases, with advantages such as out-of-the-box functionality, elastic scalability, and pay-as-you-go, are currently a research hotspot in academia and industry. Currently, cloud-native databases only support “single writer and multiple readers”, that is, read-write transactions are concentrated on a single read-write node, and read-only transactions are distributed to multiple read-only nodes. This limitation restricts the system’s ability to process read-write transactions, making it difficult to meet the demands of write-intensive businesses. To this end, this study proposes the D3C (deterministic concurrency control cloud-native database) architecture. It breaks through the limitation of “single writer and multiple readers” and supports concurrency execution of read-write transactions on multiple read-write nodes by designing a cloud-native database transaction processing mechanism based on deterministic concurrency control. D3C splits transactions into sub-transactions and independently executes them on each node according to a predefined global order, ensuring serializability for transaction execution on multiple read-write nodes. Additionally, this study introduces mechanisms like asynchronous batch data persistence mechanisms based on multi-version to ensure transaction processing performance and proposes a consistency point-based fault recovery mechanism to achieve high availability. Experimental results show that D3C can achieve 5.1 times the performance of the “single writer and multiple readers” architecture in write-intensive scenarios while meeting the key requirements of cloud-native databases.

Abstract:Databases are important foundational components in computer services. However, performance anomalies may occur during their operation, affecting business service quality. How to diagnose performance anomalies in databases has become a hot issue in industry and academia. Recently, a series of automated database anomaly diagnosis methods have been successively proposed. They analyze the runtime status of the database and determine the overall database anomaly types. However, with the continuous expansion of data scale, distributed databases are becoming an increasingly popular solution in the industry. In a distributed database, which is composed of multiple nodes, existing anomaly diagnosis methods struggle to effectively locate node anomalies, fail to identify compound anomalies across multiple nodes, and are unable to perceive the complex performance influence relationships between nodes, lacking effective diagnostic capabilities. To address these challenges, this study proposes a distributed database diagnosis method for compound anomalies, named DistDiagnosis. It models the anomalous state of distributed databases using a Compound Anomaly Graph, which not only represents anomalies at each node but also effectively captures the correlations between nodes. DistDiagnosis introduces a node correlation-aware root cause anomaly ranking method, effectively locating root cause anomalies according to the influence of nodes on the database. In this study, anomaly testing cases for various scenarios are constructed on OceanBase, a domestically developed distributed database. Experimental results show that DistDiagnosis outperforms other advanced baselines, achieving the AC@1, AC@3, and AC@5 values of 0.97, 0.98, and 0.98. Compared to the second-best method, DistDiagnosis improves accuracy by up to 5.20%, 5.45%, and 4.46% in each diagnostic scenario.

Abstract:Cloud-native databases leverage cloud infrastructure to provide highly available and elastically scalable data management, and they have experienced rapid development in recent years. As a transparent, tamper-proof, and traceable database system, blockchain sharding is the most direct and promising solution for scaling up blockchain systems. By taking advantage of the elastic scalability of cloud infrastructure, more flexible scaling can be achieved. This study first summarizes three key technical challenges addressed by current blockchain sharding: the security of node partitioning, efficient on-chain data sharding, and cross-shard transaction processing. It reviews the research status of these three issues, introduces and compares the corresponding solutions for each issue, and also discusses the new challenges these solutions face in cloud-native environments. Then, around these three dimensions, a comprehensive analysis and comparison of all solutions are conducted from the perspective of the overall impact on the blockchain system. Finally, the paper analyzes the development trends in blockchain sharding technology and points out several research directions that deserve further exploration.

Abstract:Single-master multi-slave is the mainstream architecture of cloud-native databases. In the cluster, slave nodes can share the read-only requests of the master node, while write requests are handled by the master node. Based on this, to further meet the demands of large-scale transaction expansion, some cloud databases attempt to implement multi-write transaction expansion. One possible approach to multi-write expansion is to introduce shared cache among computing nodes to support cross-node data access. For shared-cache database systems, the overhead of cross-node remote access is significantly higher than that of local access. Therefore, the design of cache protocol is a crucial factor that affects system performance and scalability. This study proposes two innovative improvements to the coherence protocol and implements PG-RAC, a shared-cache database, which supports multi-write transactions based on PostgreSQL. On one hand, PG-RAC proposes a new distributed chained routing strategy, which disperses routing information among computing nodes. Compared to the routing strategy that utilizes single-node directory management, it reduces the average transaction latency by approximately 20%. On the other hand, this study also enhances the duplicate page invalidation mechanism by separating invalidation operations from the transaction path, reducing the latency of the critical path in the transaction. Based on this, PG-RAC takes advantage of the characteristics of multi-version concurrency control (MVCC) and further proposes to delay the invalidation point of duplicate pages, which effectively improves cache utilization. TPC-C experimental results show that for a cluster with 4 compute nodes, the throughput is nearly 2 times that of PostgreSQL and 1.5 times that of the distributed database Citus.

Abstract:Secure computation of federated multi-party databases can perform federated querying or federated modeling on private data from multiple databases while preserving data privacy. Such a federation is typically a loosely organized group where the participating databases may dropout unexpectedly. However, existing multi-party secure computation systems usually employ privacy-preserving computation schemes like secret sharing, which require participants to remain online, resulting in poor system availability. Moreover, these systems are unable to predict the number of users or request rates when providing services externally. If the system is deployed on a private cluster or rented virtual machines from a cloud computing platform, it will experience increased latency during sudden bursts of requests and resource waste when the request workload is low, leading to poor overall scalability of the system. With the advancement of cloud computing technology, serverless computing has emerged as a new cloud-native deployment paradigm that offers excellent elastic resource scaling. This study designs a system architecture and an indirect communication scheme within the serverless computing framework to architect a highly scalable and highly available multi-party database secure computation system. This system can tolerate database node disconnections and automatically scale system resources in response to user request traffic changes. A system prototype based on Alibaba Cloud and OceanBase database is implemented. Comprehensive experimental comparisons are conducted. The results show that the proposed system outperforms existing systems in terms of computational cost, system performance, and scalability for tasks such as low-frequency queries and horizontal modeling. It can save up to 78% in computational costs and improve system performance by over 1.6 times. The shortcomings of the proposed system for tasks such as complex queries and vertical modeling are analyzed.

Abstract:Many computational problems on graphs are NP hard, so a natural strategy is to restrict them to some special graphs. This approach has seen many successes in the last few decades, and many efficient algorithms have been designed for problems on graph classes including graphs of bounded degree, bounded tree-width, and planar graphs, to name a few. As a matter of fact, many such algorithmic results can be understood in the framework of the so-called algorithmic meta-theorems. They are general results that provide efficient algorithms for decision problems of logic properties on structural graphs, which are also known as model-checking problems. Most existing algorithmic meta-theorems rely on modern structural graph theory, and they are often concerned with fixed-parameter tractable algorithms, i.e., efficient algorithms in the sense of parameterized complexity. On many well-structured graphs, the model-checking problems for some natural logics, e.g., first-order logic and monadic second-order logic, turn out to be fixed-parameter tractable. Due to varying expressive power, the tractability of the model-checking problems of those logics have huge differences as far as the underlying graph classes are concerned. Therefore, understanding the maximum graph classes that admit efficient model-checking algorithms is a central question for algorithmic meta-theorems. For example, it has been long known that efficient model-checking of first-order logic is closely related to the sparsity of input graphs. After decades of efforts, our understanding of sparse graphs are fairly complete now. So much of the current research has been focused on well-structured dense graphs, where challenging open problems are abundant. Already there are a few deep algorithmic meta-theorems proved for dense graph classes, while the research frontier is still expanding. This survey aims to give an overview of the whole area in order to provide impetus of the research of algorithmic meta-theorem in China.

Abstract:As software vulnerabilities grow in type, volume, and complexity, researchers have proposed various techniques to help developers discover, detect, and localize vulnerabilities. However, researchers still need to exert considerable effort to manually repair these vulnerabilities. In recent years, some researchers have focused on automated software vulnerability repair. However, such a task is merely considered a generic text generation problem by the current advanced technology, and the detects are not located. As a result, the generation space of the repair program is large, and the generated repair program is low-quality. Providing developers with such low-quality repairs affects the efficiency and effectiveness of vulnerability repair. To solve the above problems, a general type vulnerability repair approach based on chain-of-thought is proposed in this study, which is named CotRepair. By utilizing the chain-of-thought technology, the model first predicts the locations that are most likely to contain vulnerable code, and then generates the repair program more accurately based on the predicted locations. The experimental results show that CotRepair outperforms the baselines in various metrics, and the effectiveness of the proposed approach is demonstrated from multiple aspects.

Abstract:With the development of deep learning technologies such as pre-trained models, represented by Transformer, large language models (LLMs) have shown excellent comprehension and creativity. They not only have an important impact on downstream tasks such as abstractive summarization, dialogue generation, machine translation, and data-to-text generation but also exhibit promising applications in multimodal fields such as image description and visual narratives. While LLMs have significant advantages in performance, deep learning-based LLMs are susceptible to hallucinations, which may reduce the system performance and even seriously affect the trustworthiness and broad applications of LLMs. The accompanying legal and ethical risks have become the main obstacles to their further development and implementation. Therefore, this survey provides an extensive investigation and technical review of the hallucinations in LLMs. Firstly, the hallucinations in LLMs are systematically summarized, and their origin and causes are analyzed. Secondly, a systematical overview of hallucination evaluation and mitigation is provided, in which the evaluation and mitigation methods are categorized and thoroughly compared for different tasks. Finally, the future challenges and research directions of the hallucinations in LLMs are discussed from the perspectives of evaluation and mitigation.

Abstract:In the era of big data, the sample scale and the dynamic update and variation of dimensionality greatly increase the computational burden. Most of these data sets do not exist in the form of a single data type but are more often hybrid data containing both symbolic and numerical data. For this reason, scholars have proposed many feature selection algorithms for hybrid data. However, most of the existing algorithms are only applicable to static data or small-scale incremental data and cannot handle large-scale dynamic changing data, especially large-scale incremental data sets with changing data distribution. To address this limitation, this paper proposes a multi-granulation incremental feature selection algorithm for dynamic hybrid data based on an information fusion mechanism by analyzing the variations and updates of granularity space and granularity structure in dynamic data. The algorithm focuses on the mechanism of granularity space construction in dynamic hybrid data, the mechanism of dynamic update of multiple data granularity structures, and the mechanism of information fusion for data distribution variations. Finally, the paper verifies the feasibility and efficiency of the proposed algorithm by comparing the experimental results with other algorithms on the UCI dataset.

Abstract:Recently, graph convolutional network (GCN), as a powerful graph embedding technology, has been widely applied in the field of recommendation. The main reason is that most of the information in recommender systems can be modeled as graph-structured data, and GCN, as a deep learning model that operates on graph structures, helps to explore the potential interactions between users and items in graph-structured data, to enhance the performance of the recommender systems. Since the modeling of recommender systems usually needs to collect and process a large amount of sensitive data, it may face the risk of privacy leakage. Differential privacy, as a privacy protection model with a solid theoretical foundation, has been widely used in recommender systems to solve the problem of personal privacy leakage. Currently, the research based on differential privacy is mainly oriented to independent and identically distributed data models. However, data within GCN-based recommender systems is highly correlated and not independent, making the existing privacy protection methods less effective. To solve the problem, this study proposes a graph convolutional collaborative filtering recommendation algorithm based on Rényi differential privacy (RDP-GCF for short), aiming to achieve a balance between privacy protection and utility while ensuring the security ofuser-item interaction data. The algorithm first utilizes GCN techniques to learn the embedding vectors for users and items. Then, the Gaussian mechanism is used to randomize the embedding vectors, and a sampling-based method is used to amplify the privacy budget and minimize the injection of differential noise, thereby improving the performance of the recommender system. Lastly, the final embedding vectors of the users and items are obtained by a weighted fusion and applied to the recommendation tasks. The proposed algorithm is validated through experiments on three publicly available datasets. The results show that compared to existing similar methods, the proposed algorithm more effectively achieves a balance between privacy protection and data utility.

Abstract:Pre-training knowledge graph (KG) models facilitate various downstream tasks in e-commerce applications. However, large-scale social KGs are highly dynamic, and the pre-training models need to be updated regularly to reflect the changes in node features caused by user interactions. This study proposes an efficient incremental update framework for the pre-training KG models. The framework mainly includes a bidirectional imitation distillation method to fully use the different types of facts in new data, and a sampling strategy based on samples’ normality and abnormality is proposed to sample the most valuable facts from all new facts to reduce the training data size, and a reverse replay mechanism is proposed to generate high-quality negative facts that are more suitable for the incremental training of social KGs in e-commerce. Experimental results on real-world e-commerce datasets and related downstream tasks demonstrate that the proposed framework can incrementally update the pre-training KG models more effectively and efficiently compared to state-of-the-art methods.

Abstract:Benefiting from the rapid development of information technology and the widespread adoption of medical information systems, a vast amount of medical knowledge has been accumulated in medical databases, including patient clinical treatment events and medical expert consensus. It is crucial to extract knowledge from these medical facts and effectively manage and utilize them, which can advance the automation and intelligence of diagnosis and treatment. Knowledge graphs, as a novel knowledge representation tool, can effectively mine and organize information from abundant medical facts and have received extensive attention in the medical field. However, existing medical knowledge graphs often suffer from limitations such as small scale, numerous restrictions, poor scalability, and so on, leading to a limited ability to express knowledge from medical facts. To address these issues, this proposes a bilayer medical knowledge graph architecture and employs information extraction techniques on both English patient clinical treatment events and Chinese medical expert consensus to construct a billion-scale medical knowledge graph that is cross-lingual, multimodal, dynamically updated, and highly scalable, aiming to provide more accurate, intelligent medical services.

Abstract:Existing multi-view attributed graph clustering methods usually learn consistent information and complementary information in a unified representation of multiple views. However, not only will the specific information of the original views be lost under the method of learning after fusion, but also the consistency and complementarity are difficult to balance under the unified representation. To retain the original information of each view, this study adopts the method of learning first and then fusing. Firstly, the shared representation and specific representation of each view are learned separately before fusion, and the consistent information and complementary information of multiple views are learned more fine-grained. A multi-view attributed graph clustering model based on shared and specific representation (MSAGC) is constructed. Specifically, the primary representation of each view is obtained by a multi-view graph encoder, and then the shared information and specific information of each view are obtained. Then the consistent information of multiple views is learned by aligning the view shared information, the complementary information of multiple views is utilized by combining the view specific information, and the redundant information is processed through the difference constraint. After that, the topological structure and attribute feature matrix of the multi-view decoder reconstruction graph are trained. Finally, the additional self-supervised clustering module makes the learning and clustering tasks of graph representation tend to be consistent. The effectiveness of MSAGC is well verified on real multi-view attributed graph datasets.

Abstract:Keyword-based auditing (KA) technology is a crucial measure to achieve cost-effectiveness in cloud auditing applications. Different from probabilistic auditing, which verifies outsourced data by random sampling and verification, KA considers the auditing requirements of multi-user and multi-attribute data by performing keyword searches and targeted audits. KA can significantly reduce auditing costs. However, existing KA schemes usually focus only on auditing the efficiency of target data while paying little attention to remedial measures such as fault localization and data recovery after audit failures. This lack of attention to remediation measures does not guarantee data availability. Therefore, this study proposes a keyword-based multi-cloud auditing scheme (referred to as KMCA) that leverages smart contracts to enable targeted auditing, batch fault localization, and data recovery. Specifically, the targeted auditing module defines the keyword-file mapping based on the searchable encryption index structure and employs Bloom filters’ false-positive rate characteristic to hide keyword frequency and protect privacy. The fault localization module uses a binary search approach to locate error-prone cloud servers in batches and fine-grained localization of corrupted data. The data recovery module formulates multi-cloud redundant storage and data recovery strategies to avoid single-point failure and improve storage fault tolerance. Under the random oracle model, KMCA is provably secure. Performance analysis shows that KMCA is feasible.

Abstract:In recent years, online transactions of digital collections have been increasing, with platforms such as Alibaba Auctions and OpenSea facilitating their circulation in the market. However, the bidder’s bidding privacy is at risk of being disclosed during an online auction. To address this issue, this study proposes a privacy-preserving online auction approach based on the homomorphic property of SM2, which not only protects the users’ bidding privacy but also ensures the usability of the bidding data. Specifically, this study creates a homomorphic encryption scheme based on SM2, encrypting bidders’ bidding information and constructing a piece of noisy bidding information to conceal the privacy data. The efficiency of the online auction privacy preservation approach is improved by integrating the Chinese reminder theorem and baby step giant step (CRT-BSGS) into the homomorphic encryption process with SM2, which has proved to be more efficient than the Paillier algorithm. Finally, the security and efficiency of the proposed scheme are verified in detail.

Abstract:The assessment of adversarial robustness requires a complete and accurate evaluation of deep learning models’ noise resistance by combining the attack ability and noise magnitude of adversarial samples. However, the lack of completeness in the adversarial robustness evaluation metric system is a key problem with the existing adversarial attack and defense methods. The existing work on adversarial robustness evaluation lacks analysis and comparison of the evaluation metric system. The impact of attack success rate and different norms on the completeness of the robustness evaluation metric system and the restrictions on designing attack and defense methods are neglected. In this study, the adversarial robustness evaluation metric system is discussed in two dimensions: norm selection and metric indicators. The theoretical analysis of robustness evaluation completeness is carried out from three aspects: the inclusion relation of the evaluation metric domain, robustness description granularity, and the order relationship of the robustness evaluation metric system. The following conclusions are drawn: using noise statistical quantities such as the mean results in a larger and more comprehensive definition domain of evaluation indicators compared to using attack success rates, while also ensuring that any two adversarial sample sets can be compared. Using the $L_2 $ norm is more complete in the description of adversarial robustness evaluation compared to using other norms. Extensive experiments on 23 models and 20 adversarial attacks across 6 datasets validate these conclusions.

Abstract:The cross-shard state transition protocol is the basis for ensuring the atomicity of cross-shard transactions, and its efficiency directly affects the performance of the sharding system. The cross-transaction process of the existing protocols can be divided into three phases: source-shard state move-out, cross-shard state transition, and destination-shard state move-in. These phases are executed sequentially, and all phases are tightly coupled. This paper proposes the ChannelLink cross-shard state transition protocol based on the off-chain state channel. Since the off-chain channels are highly flexible and can be confirmed instantly, the ChannelLink protocol can effectively decouple the tightly coupled three-phase process, reducing the average cost of cross-shard transactions, and improving state transition efficiency. On this basis, this paper designs a low-overhead off-chain channel routing algorithm. This algorithm solves the optimal state routing scheme by improving the genetic algorithm based on the characteristics of state transition transactions and off-chain channel topology. It reduces the user’s cross-shard state transition overhead and guarantees transition efficiency. Finally, this paper implements the ChannelLink protocol prototype system and uses Bitcoin transactions and the Lightning Network state to construct the dataset for experimental verification. Results show that in a scenario with 16 shards and a cross-shard transaction ratio of 5.21%, the sharding system integrated with the ChannelLink protocol can improve the throughput by 7.04%, reduce the transaction confirmation latency by 52.51%, and reduce the cost of cross-shard state transition by more than 45.44%. Meanwhile, the performance advantages of the ChannelLink protocol gradually increase as the number of shards and the cross-shard transaction ratio increase.

Abstract:Federated learning, a framework for training global machine learning models through distributed iterative collaboration without sharing private data, has gained prevalence. FedProto, a widely used federated learning approach, employs abstract class prototypes, termed feature maps, to enhance model convergence speed and generalization capacity. However, this approach overlooks the verification of the aggregated feature maps’ accuracy, risking model training failures due to incorrect feature maps. This study investigates a feature map poisoning attack on FedProto, revealing that malicious actors can degrade inference accuracy by up to 81.72% through tampering with the training data labels. To counter such attacks, we propose a dual defense mechanism utilizing knowledge distillation and feature map validation. Experimental results on authentic datasets demonstrate that this defense strategy can enhance the compromised model inference accuracy by a factor of 1 to 5, with only a marginal 2% increase in operational time.

Abstract:The scene sketch is made up of multiple foreground and background objects, which can directly and generally express complex semantic information. It has a wide range of practical applications in real life and has gradually become one of the research hotspots in the field of computer vision and human-computer interaction. As the basic task of the semantic understanding of scene sketch, scene sketch semantic segmentation is rarely studied. Most of the existing methods are improved from the semantic segmentation of natural images, which cannot overcome the sparsity and abstraction of sketches. To solve the above problems, this study proposes a graph Transformer model directly from sketch strokes. The model combines the temporal-spatial information of sketch strokes to solve the semantic segmentation task of free-hand scene sketches. First, the vector scene sketch is constructed into a graph with strokes as the nodes of the graph and temporal and spatial correlations between strokes as the edges of the graph. The temporal-spatial global context information of the strokes is then captured by the edge-enhanced Transformer module. Finally, the encoded temporal-spatial features are optimized for multi-classification learning. The experimental results on the SFSD scene sketch dataset show that the proposed method can effectively segment scene sketches using stroke temporal-spatial information and achieve excellent performance.

Abstract:Deep learning has yielded remarkable achievements in many computer vision tasks. However, deep neural networks typically require a large amount of training data to prevent overfitting. In practical applications, labeled data may be extremely limited. Thus, data augmentation has become an effective way to enhance the adequacy and diversity of training data and is also a necessary link for the successful application of deep learning models to image data. This study systematically reviews different image data augmentation methods and proposes a new classification method to provide a fresh perspective for studying image data augmentation. The advantages and limitations of various data augmentation methods are introduced from different categories, and the solution ideas and application values of these methods are elaborated. In addition, commonly used public datasets and performance evaluation indicators in three typical computer vision tasks of semantic segmentation, image classification, and object detection are presented. Experimental comparative analysis of data augmentation methods is conducted on these three tasks. Finally, the challenges and future development trends currently faced by data augmentation are discussed.